Deep Learning for Woodland Soil Biodiversity Identification and Monitoring

Soils are one of the most important reservoirs of biodiversity [1], and soil fauna is a critical component of functioning ecosystems across the world. Soil fauna participates in the maintenance of soil structure, as well as playing important roles in carbon cycling and decomposition, nutrient availability and food production, water flow and pest regulation amongst other services [2]. The sampling of soil fauna follows standardised collection methods, which have been well-developed for a wide range of taxa [3]. The remaining challenge is the availability of taxonomical expertise due to species-rich communities ranging from microscopic to macroscopic size [4], particularly for obscure or historically difficult groups, and the time involved in sample preparation and identification. An experienced soil fauna taxonomist can spend multiple days per soil sample, with research projects often generating hundreds of such samples to obtain an accurate representation of soil communities. The use of novel technologies such as deep learning (DL) methods are conducive to processing data in an automated and a time-efficient manner, allowing to process samples in a fraction of the time currently required by experts.

The novel open access ‘CollembolAI’ workflow developed by Sys et al. [5], and similar endeavours on their way across the world, provide an effective workflow to automate soil fauna identification using DL. This approach frees resources for the establishment of cost-efficient long-term monitoring programs that can address taxonomic, spatio-temporal and resolution gaps in biodiversity knowledge, allowing for a faster response to faunal changes in a conservation setting due to sudden changes on ecosystem condition. Further, through the development of soil fauna identification DL methods, we can efficiently add value to existing DL projects (such as camera-trapping, and aerial woodland images) by establishing links between above- and belowground biodiversity. Filling knowledge gaps will inform biodiversity protection policy and the implementation of resilience management plans for woodlands and associated habitats.

Research Questions

Through implementation of a DL workflow for woodland soil specimen identification, we aim to address the following research questions:

Q1) What is the algorithm's capability for discerning individual specimens from soil fauna community samples, and successful characterisation of targeted taxa (i.e., springtails and mites).

Q2) Woodland soil specimens are likely to contain multiple species with some classes underrepresented in annotated data. What is the role of self-supervision, one-shot and few-shot learning approaches to increase taxonomic performance on complex samples with multiple communities.

Q3) What level of taxonomic identification can be achieved between taxa and can the digital area (i.e. pixels) of specimens serve as a proxy for specimen biomass [6].

Q4) What is the effect of a lower taxonomic resolution than species, e.g., family for downstream soil characterisation.

Objectives

In this cross-disciplinary project, the PhD candidate will work closely with FR’s soil fauna taxonomist to address three objectives over the course of the project:

O1) We will apply the novel, proof-of-concept workflow to soil fauna community samples which have already had specimens fully taxonomically identified and biomass estimated by soil taxonomic specialists, from past and ongoing forest, peatland and agroforestry soil research projects across the UK.

O2) We will evaluate workflow performance based on image resolution, as well as model evaluation metrics such as accuracy, precision, and recall.

O3) We will assess data processing time and dataset size required to train the model and assess the suitability of this approach for the identification of microscopic fauna.